Machines talking to machines

Automated screening reads applications that were partly automated. Asynchronous video and questionnaire tools already assess candidates who prepared their answers with a model. In that sense the loop is closed at the top of the funnel.

What is being described as "AI interviewing AI" is mostly this: two automated layers exchanging documents, with no human judgement on either side until later.

Neither side chose it deliberately. Employers automated because a popular role attracted more applications than anyone could read, candidates automated because a keyword filter can be satisfied by supplying keywords, and each response was rational given the other.

Hiring exists to answer whether a specific person can do specific work and function in a specific team. An exchange between two systems answers neither, however sophisticated the exchange.

If candidates delegate answering and employers delegate assessing, the process generates a lot of activity and no information. Both sides notice — employers because their hires do not improve, candidates because they are rejected without ever reaching a person.

Accountability is the harder obstacle. Someone signs off on a hire and answers for it eighteen months later, and that person will want to have met the candidate — which is a structural reason rather than a sentimental one, and it does not weaken as the technology improves.

When a signal becomes automatable, its value falls and hiring moves toward signals that cannot be. That has happened before with keyword-optimised CVs and with formulaic cover letters.

The plausible direction is fewer, higher-quality human interactions supported by better preparation, alongside more work-sample assessment where the output is the evidence.

Some employers are taking a counter-intuitive route that fits the same logic: asking for far less. A very short application makes a small work sample the real filter, which is cheaper to evaluate at scale than a long document nobody trusts and produces a signal candidate automation cannot fake.

• Work samples and paid trial tasks, where output is checkable
• Fewer but longer conversations with actual team members
• More weight on referrals and demonstrable track record
• Less reliance on documents as a screening signal

It is worth naming the cost rather than treating this as a symmetric arms race. Candidates rejected by an automated screen that read an automated application never reach a person who could have noticed something the system missed.

That lands hardest on the people least served by document screening already: career changers whose history does not match the target, people with unconventional backgrounds, and anyone whose strongest evidence is what they can do rather than what they can describe.

A work-sample shift helps some of those candidates and not all of them. It rewards demonstrable ability and still requires unpaid time, which is its own filter — so "more work samples" is an improvement rather than a fix.

It is good at what recruiters do badly at volume: reading every application rather than the first fifty, applying criteria consistently, responding to everyone, scheduling without a fortnight of email.

Those are real improvements for candidates as well as employers, and they do not require automating the judgement. The distinction between administering a process and deciding it is the one worth holding.

Most complaints about recruiting are capacity failures rather than indifference — no response, slow scheduling, no feedback — and capacity problems have technical solutions in a way motivation problems do not.

Automation will keep expanding in screening and administration, and the decision will stay human — not out of sentiment, but because an employer accountable for a hire will want to have met the person.

Where that breaks down, it produces bad hiring rather than efficient hiring, and organisations correct it once they notice. The equilibrium is automated logistics, human judgement.

Regulation is the other force, and it points the same way. Several jurisdictions already require bias auditing, candidate notice or a route to human review for automated hiring decisions, which makes a fully machine-decided process a legal problem as well as a bad one.

For a candidate the practical reading is unchanged. Get to a person as early as possible, and make the evidence that survives automation — work, referrals, specifics — the part you invest in.